NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters that activates 12 billion per token. It handles reasoning, instruction following, tool use, and text, image, document and video understanding in one package, so teams don't have to run a separate vision-language model next to a text model and shuttle context between them.

The company says it delivers up to 5x higher throughput on agentic tasks on a single GPU versus open models its size, and it ships with base, post-trained and NVFP4-quantized checkpoints, training recipes, reinforcement-learning environments, and open datasets on Hugging Face — including a roughly 2.8 trillion-token Indic-language corpus built with IIT Madras and AI4Bharat.

Early customer numbers are the hook: NVIDIA says Scale AI's post-trained version for vendor credit processing at Skydance will cut inference cost 40x and latency 5x, Ramp expects up to 50% higher throughput and 10x lower latency on receipt extraction, and AI21 improved accuracy 40% on threat-intelligence work for Alice.

We think this is the more honest shape of the enterprise agent story. Most real work isn't a chatbot conversation; it's a PDF with a chart on page four, a receipt photo, a policy document and three tools to call. One open model that sees all of it, and that a company can train on its own procedures, is genuinely useful.

The open part matters to us. Base checkpoints, recipes, RL environments and datasets — including a large Indic corpus — mean a bank or hospital can adapt and run this where its data already lives instead of piping everything to someone else's API. That's a real shift in who holds the leverage.

Now the caveats. The headline customer gains are vendor-reported, and some are framed as what costs "will" do or what a partner "expects." Benchmarks against NVIDIA's own previous models tell us about progress inside the family, not necessarily against the best closed systems on a given company's messy workload. We'd treat 40x as a ceiling seen in a tuned case, not a forecast.

There's also a strategic read. Every efficiency trick here — NVFP4 on Blackwell, speculative decoding, Dynamo routing — makes the model cheapest on NVIDIA hardware. Open weights, yes; neutral stack, not quite. That's fine, but buyers should know which lock they're choosing.

The downside if this scales is the one we keep flagging with agents: a model that reads documents, triages vulnerabilities and calls tools is a model that can be fed poisoned documents and act on them. Post-training on company policy helps it follow the rules; it doesn't make it immune to someone rewriting the input. I'm watching for whether the customization tooling ships with evaluation for that, not just for accuracy.

NVIDIA positions Super VL in a tiered family: smaller Nemotron 3.5 Lightning for high-volume text, Super VL for specialized multistep work, and Nemotron 3 Ultra — from which Super VL was distilled — for complex planning.